Semantic Relation-Guided Contrastive Learning for Few-Shot Object Detection

Few-shot object detection (FSOD) aims to achieve effective detection of novel categories using only limited annotated samples. Current mainstream approaches predominantly employ transfer learning strategies, conducting base training on abundant base-class samples followed by fine-tuning with sparse novel-class samples. Recent studies demonstrate that incorporating contrastive learning can significantly enhance object classification performance in detection tasks. However, conventional contrastive learning methods typically rely on simplistic instance-level comparisons, which fail to capture inter-sample semantic relationships while neglecting quality variations among candidate detection boxes, ultimately limiting the model's generalization capability in few-shot scenarios. To address these limitations, this paper proposes a supervised contrastive learning framework based on semantic relation reasoning. First, the method adaptively selects objects with high Intersection-over-Union (IoU) scores as high-quality positive samples to optimize contrastive loss computation, thereby enhancing feature discriminability. Second, we design a reweighting mechanism for feature-augmented multi-source samples to balance their relative contributions, further improving model robustness. Experimental results demonstrate that our approach achieves competitive performance across multiple FSOD benchmark datasets, effectively validating both the method's efficacy and generalization capability.